JPH0830728A - Binarization device for image - Google Patents

Abstract

PURPOSE:To perform highly accurate binarization without being influenced by noise and contrast, etc. CONSTITUTION:This device is provided with a source image input part 10 for image-picking up images provided with a recognition processing object a1 and converting the image picked-up analog image data to source image data (b) provided with gradation, a threshold value selection part 12 for calculating a threshold value (c) for correction for the binarization of the source image data (b), an input data preparation part 14 for preparing input data X1 to Xn by correcting the density level of the source image data (b) based on the threshold value (b) for the correction and normalizing the source image data (b) and a neural network processing part 16 for binarizing the input data X1 to Xn prepared by the input data preparation part 14 by a neural network processing and converting them to output data x1 to xn.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image binarization device used for preprocessing for recognizing characters, figures, symbols, colors, etc. in an input image. The present invention relates to an image binarization device for converting into a two-tone image of "0" or "1".

This device is applied to a binarization device for binarizing a grayscale image as a preprocessing for recognizing a character in an input image in a character recognition device for optically reading a character.

[0003]

2. Description of the Related Art A conventional image binarization device is an A / D converter which performs binarization based on a threshold value. In addition, image processing has also been performed in which an image with gradation is binarized based on a threshold value. Various determination methods are known for these threshold values. For example, a density histogram of an analog image or a digital image with gradation is created, and the density value of the valley of the histogram is used as the threshold value.

[0004]

However, in such a conventional technique, first, it is difficult to select a threshold value for properly separating the background portion and the recognition target portion in the analog image, and the threshold value is automatically set. In the method of selecting the value, the accuracy does not always reach the accuracy required by the recognition means. Further, when the original image contains noise and the density value of the noise is equal to or higher than the threshold value, the noise is binarized as an effective part (a part of the recognition target such as characters). There was an inconvenience. As described above, in the conventional technique, it is difficult to remove noise and perform binarization with high accuracy.

Japanese Patent Application No. 6-85519 by the same applicant who attempts to solve this inconvenience proposes a method of binarizing a grayscale image using a recurrent neural network. In this method, the difference between the average density value of the outermost peripheral portion of the original image and each pixel is taken as preprocessing for image recognition, and the difference image is input to the recurrent neural network. Then, since the density level of the background portion could be made constant, it was possible to perform binarization while removing noise by the neural network processing.

However, this method has a disadvantage that the accuracy of binarization cannot be improved beyond a certain level. In particular, when an image with high contrast or an image with extremely low contrast is targeted, highly accurate binarization cannot be performed. That is, when an image with high contrast is targeted, this method has a disadvantage that the reference value of the neural network processing is not set to an appropriate value, and thus the image cannot be binarized with high accuracy.

On the other hand, an image with extremely low contrast is
For example, this corresponds to an original image obtained by imaging a stamped character stamped on a metal surface such as an engine block. As a pre-processing for such automatic recognition processing of the engraved character, there is a disadvantage that the image of the engraved character having extremely low contrast cannot be binarized with the accuracy required by the recognition unit.

[0008]

SUMMARY OF THE INVENTION It is an object of the present invention to provide an image binarization apparatus which is capable of improving the disadvantages of the conventional example, and in particular binarizing it with high accuracy without being affected by noise, contrast and the like. Is the purpose.

[0009]

Therefore, in the present invention, an original image input unit for capturing an image including a recognition processing target and converting the captured analog image data into original image data having gradation, and the original image input unit. A threshold value selection unit that calculates a correction threshold value for binarization of an image, and an input by correcting the density level of the original image based on the correction threshold value and normalizing the original image. The configuration is such that an input data creating unit for creating data and a neural network processing unit for binarizing and outputting the input data created by the input data creating unit by a neural network process are adopted. This aims to achieve the above-mentioned object.

[0010]

During the operation of the apparatus, the original image input section first captures an image including the recognition processing target. This is performed by generating an analog image that is a waveform of a voltage change obtained by photoelectrically converting the reflected light from the imaging target. Then, the original image input section
The analog image is converted into original image data having gradation. The threshold selection unit calculates a correction threshold for binarizing the original image data. Next, the input data creation unit corrects the density level of the original image data based on the correction threshold value. Therefore, the density level of the original image is corrected based on the correction threshold value regardless of the density of the analog image captured by the original image input unit and regardless of the density distribution.

Further, the input data creating unit creates the input data X1 to Xn by normalizing the original image data b whose density level has been corrected. Subsequently, this input data is binarized by the neural network processing unit. The binarized image data is output to a character recognition device or the like.

At this time, in the neural network processing section, the neural network is pre-learned so that the input data can be accurately recognized and binarized. For example, even if the pixel exceeds the threshold value, learning is performed so as to determine that the pixel is noise based on the relationship with the surroundings of the pixel. Therefore, in the neural network processing unit,
The noise contained in the input data is recognized and treated in the same way as the background.

[0013]

An embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of an image binarization apparatus according to the present invention. The image binarization apparatus captures an image including a recognition processing target a1 and converts the captured analog image data into original image data b with gradation, and an original image input unit 10 for converting the original image data b. Threshold selection unit 12 that calculates a correction threshold c for binarization
And an input data creation unit 14 that creates the input data X1 to Xn by correcting the density level of the original image data b based on the correction threshold value c and normalizing the original image data b. The input data creating unit 14 is provided with a neural network processing unit 16 for binarizing the input data X1 to Xn by a neural network process and converting the input data X1 to Xn into output data x1 to xn. Further, the neural network processing unit 16 outputs the output data x1
A binarized image synthesizing unit 18 for generating and outputting a binarized image by synthesizing .about.xn is additionally provided.

This will be described in detail. Original image input unit 10
Is an image scanner 10A such as an image scanner or a CCD camera, and an analog image captured by the image capture unit 10A is converted into digital image data (original image) b with gradation A / D.
It is composed of an A / D converter 10B for conversion and the character a1
An image including a recognition processing target such as is photoelectrically converted and output as a digital value. Here, first, the CCD camera 10A photoelectrically converts the reflected light of the image including the recognition processing target and outputs an analog image which is a continuous voltage change amount. Next, the A / D converter 10B converts the analog image into a digital image (original image) b having 256 gradations, for example.

The threshold selection unit 12 selects a threshold c for binarizing the original image b. This correction threshold value c
Is used to keep the density level of the original image b constant. Therefore, any method can be used as long as it can automatically obtain the rough threshold value c regardless of the type of the original image b. In this embodiment, the method conventionally used is adopted. There are various methods for selecting the correction threshold value c, such as the P-tile method and the discrimination threshold value selecting method, and the following method that works well depending on the character of the recognition target is adopted.

P-tile method: When the part occupied by the recognition target in the analog image is known in advance, this P-tile method may be used. This method is a method of calculating a density value in which the cumulative distribution from the smaller density value is P percent by using a density histogram and setting it as a threshold value c.

Modal method: When a high-contrast image is targeted, the density histogram exhibits bimodality,
This is a method of calculating the valley between the mountain of the background portion and the mountain of the target portion and setting it as the threshold value c. However, if the histogram is not smooth or if the valleys are not noticeable,
When the contrast is low, it is difficult to automatically determine the threshold value c. Therefore, in this embodiment, the following discrimination threshold selection method is adopted. As a method for automatically determining the threshold value c based on the modal method, a bimodal histogram is approximated by a least square error with the sum of normal distribution functions,
There is a method of calculating the threshold value c from estimated values such as average and variance.

Discriminant threshold selection method: This is based on the method of discriminant analysis in multivariate analysis, and the original image b for which the threshold value c is selected is classified into classes, and from the relationship between the interclass variance and the total variance. The value that best discriminates the target is selected. That is, the threshold value c (t) that maximizes the class separation degree η (t) is calculated. This introduces the class separability represented by the following equation (1) used in the discriminant analysis in order to measure the quality of the threshold position when binarized by the threshold value c (t).

[0019]

[Equation 1]

Here, σT ² (t) and σB ² (t) represent interclass dispersion and total dispersion, respectively. Then, a threshold value c (t) that maximizes η (t) is obtained. The class separation degree η (t) with respect to the threshold value c (t) can be interpreted as representing the degree to which the original image b is binary. This method works as the above modal method when the histogram is bimodal,
It has a feature that the threshold value c is automatically determined even when there is no mode and therefore it is difficult to determine the valley.

These techniques are publicly known, and are disclosed, for example, in "Image Processing Handbook" edited by the Image Processing Handbook Editorial Committee, May 20, 1992, first edition, fifth printing, pp. 278 to 280. For details on the threshold value calculation by the discriminant analysis method, see Nobuyuki Otsu, "Automatic Threshold Value Selection Method Based on Discriminant and Least Squares Criterion", IEICE Transactions,
See Vol. J63-D No. 4, pp.349-356 (April 1980). Here, a method of automatically calculating a threshold value by a simple procedure based only on the 0th and 1st order cumulative moments of the density histogram is disclosed.

The threshold selection unit 12 calculates the correction threshold c of the original image b based on these methods. In this embodiment, in order to provide a constraint on the processing time for binarization and a versatile binarization method, edge information is used to select the correction threshold value c. It does not adopt a method of calculating values or a dynamic threshold selection method. However, according to the improvement of the performance of the arithmetic processing unit, the second-order differentiation (Laplacian) is performed as the preprocessing of the discrimination threshold selection method.
A method such as using the histogram method may be adopted. In this embodiment, the main technical idea for solving the problem is to correct the density level of the original image b based on the correction threshold value c obtained by some method and then input the corrected density level to the neural network. , The correction threshold value c can be obtained, and therefore the present invention is not limited by the selection method of the threshold value c.

It is self-evident that the pattern recognition processing technique for images, voices and the like can obtain an effect which has never been obtained by using the neural network processing. Therefore, in the neural network processing technique, the technique of input data generation and the technique of neural network learning are technical problems. That is, a method of generating input data so as to maximize the potential effect of the neural network processing and learning the neural network is
This is the main technical problem of this embodiment.

Therefore, the input data creating section 14 corrects the density level of the original image b based on the correction threshold value c selected by the threshold value selecting section 12. There are various methods for this correction, but any method may be used as long as it is a method for correcting the density level based on the threshold value c. Here, the correction threshold value c is treated as a central value for binarization, and the density of the original image b is corrected with reference to this threshold value c. That is, in this embodiment, when the density value of the image is 256 gradations (0 to 255), the threshold value c is 1.
The density level of the original image is corrected to 28.

Here, the difference between the correction threshold value c and the center value of the gradation (for example, 128 for 256 gradations) is calculated, and all the density values are uniformly shifted by the difference. There is. Further, the portions that deviate from the gradation by adding up the differences are set to the maximum value and the minimum value, respectively, in order to ensure the binarization processing speed. of course,
Original image data b so that it does not overflow from the gradation that is handled
It may be possible to perform image processing on the image and then perform the correction.

FIG. 2 is an explanatory diagram for explaining the image processing in the input data creating unit 14 and the neural network processing unit 16. In the input data creation unit 14, by normalizing the image data whose density level has been shifted,
Input data X1 to Xn are generated. Since the density level is shifted, for example, when 256 gradations are normalized from 0.0 to 1.0, the density of the correction threshold value c is always 0. It becomes 5. Furthermore, the input data creation unit 14 has m original images b in the horizontal direction and n in the vertical direction.
Input data X1 to Xn to the neural network processing unit 16 are obtained by dividing the data into m pieces and outputting m pieces for each column in the vertical direction.

The neural network processing unit 16 inputs the input data X1 to Xn created by the input data creating unit 14.
Input layer 20 consisting of m units equal to the number of pixels
And an intermediate layer (hidden layer) 22 composed of an arbitrary number of units
And an output layer 24 having m units equal to the number of units of the input layer 20.

Further, in this embodiment, the neural network processing unit 16 is composed of a recurrent neural network. That is, the units of the input layer 20, the intermediate layer 22, and the output layer 24 are asymmetrical and recurrently coupled to each other without being restricted by the symmetric coupling or the unidirectional coupling. The recurrent neural network having such characteristics can accurately recognize input data containing noise. Note that the recurrent neural network itself is a well-known technique, for example, "measurement and control" Vo
l.30. No. 4 (April 1991) pp.296-301, etc.

In such a structure, in order to obtain the relationship between the input data X1 to Xn input to the input layer 20 and the output data x1 to xn output from the output layer 24, the coupling strength between the units is preset. Seek by learning.
That is, when certain input data is input to the input layer 20, if the input data includes noise, learning is performed so that the output data from which the noise is removed is output from the output layer 24. Alternatively, when a certain input data is input to the input layer 20, if the input data is a large lump (a set of “1”), “1” is output from the output layer 24 and the input data is a small lump. For example, the output layer 24 is trained to output “0”.

By using the learned neural network processing unit 16 as described above, when the input layer 20 is given unknown input data including, for example, the characters a1 and the noises a2 and a3, the noise a2 from the output layer 24 is obtained. Accurate output data of the character a1 with a3 removed can be obtained. Since the neural network processing unit 16 learns the shape of the pattern, correct pattern recognition is possible even when noise is mixed in the input pattern (“input data” in this embodiment).

The output data output from the neural network processing unit 16 in this way is the binarized image synthesizing unit 1
8 is combined and output as a binarized image to a character recognition device or the like.

Next, the operation of the analog image binarization apparatus will be described with reference to the flowchart of FIG. 3 and FIGS. 4 to 9.

First, the original image b is input by the original image input unit 10 (step S1). An example of the original image b is shown in FIG.
FIG. 4 is an explanatory diagram in which the characteristics of the original image b to be processed are emphasized and displayed, and the original image b is actually an image represented by a large number of gradations (density). Here, the original images A, B, and C are numbers “2”, “6”, and
It is composed of "4" and a background of numbers in which the density is hatched. Below the original images A, B, C, the analog images A1, A1, B1, B1, B1, B1, C1, C1
The graph of the density value (0-255) in a line is shown. The background density in the original image increases in the order of B, A, and C. If the neural network processing unit 16 tries to learn the original images A, B, and C as they are, the amount of information becomes enormous and it becomes difficult to converge the learning.

Therefore, in the above-mentioned Japanese Patent Application No. 6-85519, the difference image creating unit
By taking the difference between the average density value of the outermost peripheral portion and the density value of each pixel, an image of the density value shown in FIG. 5B is created. In this way, the background density values in the original images A, B, and C are kept constant. As a result, since the appropriate learning data has been obtained, the neural network processing unit 16
It was supposed to be binarized correctly.

However, this Japanese Patent Application No. 6-85519
In the issue, it was found that there is a limit to the increase in binarization accuracy due to the high and low contrast. That is, although binarization is favorably performed on an image having low contrast, noise separation was not favorably performed on an image having high contrast. That is, as shown in the density graph of the number “2” in FIG. 5B, one of the causes is that the reference value of the neural network is set lower than the density of the original image in an image with high contrast. As a result, the noise and the recognition target could not be well separated. Further, since the background level is made uniform, the background part can be favorably made constant if the density histogram is the original image data exhibiting bimodality, but an image with a small density difference between the recognition target a1 and the background part, That is, even an image with extremely low contrast could not be binarized accurately.

Therefore, in the present embodiment, the input data creation unit 14 first selects the correction threshold value c as shown in FIG. 5C (step S2), and then, FIG. )
As shown in, the density level of the original image data b is shifted based on the correction threshold value c (step S3).
Next, the input data is generated by normalizing this image (step S4). Therefore, the neural network processing unit 16 always learns using the correction threshold value c as the reference value, so that the characteristics of the original image data b can be learned satisfactorily. That is, the neural network processing unit 1
Since No. 6 self-organizes on the basis of the correction threshold value c, it is possible for itself to have a configuration that matches the characteristics of the recognition target and the processing purpose.

Further, as shown in FIG. 6, the input data creating section 14 creates the input data X1 to Xn by dividing the normalized original image b into pixels for each column (step S5).

Next, the neural network processing unit 1
Input data X1 to Xn are sequentially input to the input layer 20 of No. 6 (step S6). Then, the output data x1 to xn corresponding to the input data X1 to Xn are sequentially output from the output layer 24 (step S7). For example, as shown in FIG. 7A, input data X1 is output data x1,
Output data x2 is output for the input data X2 as shown in FIG. 7B, and output data x3 is output for the input data X3 as shown in FIG. 7C.

Finally, the binarized image synthesizing unit 18 outputs the output data x converted by the neural network processing unit 16.
1 to xn are combined and output as a binarized image d (step S8).

8 and 9 are plan views showing an example of a binarized image. 8 shows a case where the contrast of the original image b is high, and FIG. 9 shows a case where the contrast of the original image b is low. 8A and 9A are original images,
8 (B) and 9 (B) are images binarized simply by threshold values by the discrimination threshold selection method, and FIGS. 8 (C) and 9 (B).
(C) is a binarized image according to the present invention. As is clear from this figure, and as described above, according to the present invention, it is possible to obtain a highly accurate binarized image that cannot be obtained by the conventional technique.

The neural network processing unit 16
It is most preferable to use a recurrent neural network as in this embodiment, but the present invention is not limited to this, and a feedforward network or the like may be used.

The input data creating section 14 creates the input data by dividing the difference image c into n pieces in the vertical direction, but inputs the data by dividing the difference image c into m pieces in the horizontal direction. You may create the data. Further, the division is not limited to the vertical and horizontal directions, and the division may be performed diagonally, and further, not limited to one row, two or more rows may be collectively divided.

Next, an engraved character recognition device to which the present invention is applied will be described with reference to FIGS.

In the production process of automobiles, engraved characters are used as instruction symbols for assembled parts and vehicle body numbers. Unlike printed characters, engraved characters stamped on the metal surface have excellent environmental resistance and little change over time. However, the color difference with the surroundings is small, and due to the method of treating the marking surface, oil stains, scratches, etc.
It is not clear as a line drawing.

When recognizing such a stamped character by image processing, preprocessing such as image binarization, noise removal, and character cutout is required. However, since unclear character images have low contrast and a lot of noise, it is difficult for the conventional binarization method to separate the character part and the background part and accurately extract only the character part. is there.

Hitherto, as a method of binarizing printed characters, a method using a density histogram and a method using information such as an image differential value, an edge, a boundary point, and a contour line have been proposed. In the method using the density histogram,
It is difficult to perform highly accurate binarization for an image in which the contrast between the character portion and the background portion is extremely low or an image in which the area ratio between the character portion and the background portion is large. Also, the method of calculating the density histogram by adding the differential value of the image, the method of using the degree of coincidence of the boundary points and edges of the image to determine the threshold of binarization, the length of the character outline and the area of the character The method using the ratio of
There are problems such as increase in calculation time, complicated parameter setting, and difficult edge detection.

Recently, a method of applying the learning and recognition ability of a pattern by a hierarchical neural network to edge extraction and binarization of a grayscale image has been studied, but the data that can be handled by the network is 3 × 3 or 5.
Since it is limited to the local region of x5, sufficient binarization accuracy is not obtained.

On the other hand, an attempt has been made to learn and recognize time-series patterns by a recurrent neural network (hereinafter referred to as RNN) having an internal state in itself for a hierarchical neural network which performs static pattern conversion. Is made.

In this embodiment, RNN is applied to the binarization of the stamped character image by treating the stamp as a time series pattern in which the self-portrait changes in the horizontal direction or the vertical direction. Here, the RNN is made to learn the binarization of a low-contrast image and the discrimination of noise as time-series data. Further, the stamped character image is binarized by inputting it as time series data in the horizontal direction or the vertical direction into the learned RNN.

In the engine block, the numeral "1",
A three-digit character string using two characters "2" and a four-digit character string using three letters "A", "B", and "C" are engraved. In this embodiment, the number "1",
The image of "2" was used.

As shown in FIG. 10, the engraved character recognition device comprises the above-described image binarization device and a character recognition device for recognizing characters from the binarized image output by this binarization device. There is. The binarization device 42 includes an original image input unit 30 that generates an analog image by photoelectric conversion and an A / D that A / D-converts the analog image to generate an original image with 256 gradations.
The conversion unit 32 and the binarization unit 34 that binarizes the original image are provided. The original image input unit 30 here is an incandescent lamp that irradiates a marking character from a direction of 45 degrees, and a monochrome CCD that is installed in the opposite direction of the incandescent lamp and images the marking character.
Equipped with a camera.

Further, the character recognition device 44 performs character recognition by comparing the character cutout unit 36 for cutting out the character portion from the binarized image with the data output by the character cutout unit 36 and the character sample data. And a recognition unit 38. Further, the respective operation units of the binarizing device 42 and the character recognizing device 44 are controlled by the control means 40 in terms of their operation timing and data transfer.

The original image picked up by the image input unit is 640 horizontal.
[Dot] A vertical image of 400 [dots] and 256 gradations of a monochrome image, and the density of one dot is 0 to 255. In the present embodiment, characters are framed one by one in a fixed frame from this captured image, and 100 × 100 [dots]
It is handled as a character image of.

FIG. 11 shows an engraved character image string and the density distribution of one line in the image. 11 (a) and 11
(B) is an image in which the average density value of the background part is about 190, the average density value of the character part is about 240, the density difference between the background part and the character part is 50 or more, and the contrast is high and the noise is small. 11 (c) and 11 (d), the average density value of the background portion is about 230, the density difference of the character portion is about 20, the average density value of the background portion and the character portion is about 250, and the contrast is It is a low image.

In this embodiment, the image is treated as time series data and binarized. In the case of the image sequence shown in FIG. 11, the density distribution of one line shown in the graph is time-series data obtained by dividing the image into 100 in the horizontal direction.

FIG. 12 shows a flow chart of the marking character recognition processing.

To recognize the engraved character, pre-processing such as binarization, clipping, noise removal, etc. is performed on the image captured first. By preprocessing, data of only the character part is extracted from the image, and the feature value of the character is extracted for recognition.

The binarization process of the original image is performed as described above.
First, a threshold value is calculated from the input image by the discriminant analysis method (step S21). Next, the obtained threshold value is an intermediate value 128 of the density distribution (here, 0 to 255).
The density value of the entire image is shifted so as to become (step S22). That is, the obtained threshold value and density value 128
And the difference is added to the density value of each pixel of the image to uniformly shift all the density values. This is a pre-process that is performed in order to make the density value level constant irrespective of the state of the object to be imaged, the environment of the image pickup, and the like, and allows the subsequent neural network processing to operate favorably. Next, the density value from 0 to 255 of each pixel is changed from 0.0 to 1.
It is normalized to 0 and used as input data to the RNN. The created input data is input to the RNN, and the output result at that time is binarized (step S23).

Next, character recognition processing is performed. In the character recognition process, first, a character portion is cut out from the binarized image (step S25), and then the characteristic value of the character is extracted from the cut out image (step S26). Further, character recognition is performed by performing pattern recognition processing based on this feature value (step S27).

The RNN is composed of an input layer, an intermediate layer (hidden layer), and an output layer. There are the following six types of connections in this network, and all connections are in the layers of the intermediate layer and the output layer. Input layer to middle layer Input layer to output layer Middle layer to output layer Middle layer to middle layer Output layer to middle layer Output layer to output layer

Due to the limitation of the memory capacity of the computer,
In this embodiment, the number of units in each of the input layer, the intermediate layer, and the output layer is 5, and the time step of the learning data is 30. A time step is the number of pieces of data that changes in the time direction of time series data.

The operation of each unit of the RNN used in this experiment is expressed by the following equation. Here, the input layer is I, the intermediate layer is H, and the output layer is O. Further, X _i (t): external input of unit i x _i (t): internal state value of unit i y _i (t): output value of unit i w _ij : coupling load from unit j to unit i f _i : Response function of unit i Y _i (t): As a teacher output of unit i, when _i ∈ I, y _i (t) = x _i (t) = X _i (t) (2)

When i ∈ H, O,

[0064]

[Equation 2]

[0065]

(Equation 3)

The evaluation function at the time of learning is

[0067]

[Equation 4]

It is represented by

As a supervised learning rule of the RNN, a BPTT (Back Propagat) for obtaining a gradient of E (w) by using a variational method.
Ion Through Time) is used. If Lagrange's undetermined constant is λi,

[0070]

(Equation 5)

From the boundary conditions δx _i (0) = 0 and λ _i (T) = 0, the gradient of E with respect to w _ij is given by the following equation.

[0072]

(Equation 6)

Therefore, the coupling load between the units is updated by the value shown in the following equation. However, the differential equation is discretized with respect to time.

[0074]

(Equation 7)

In the conventional binarization method, a method of obtaining a threshold value and dividing each image into 0 or 1 (white or black) is adopted. In such a method, since the information on the relationship between the pixel to be binarized and the surrounding pixels is not used, the binarization is performed without knowing whether the pixel to be binarized is a character part or noise. Will be done.

In this embodiment, in order to reflect whether the pixel to be binarized is a character part or noise in the binarization, the character part is 1, the background part is 0, and the noise is 0. Is learning. Here, a lump of 3 dots or less is noise, and a lump of 4 dots or more is a character.

FIG. 13 shows an example of learning data. The upper row is the input data and the lower row is the teacher data. The vertical direction represents units, and the upper row corresponds to the units 1 to 5 in the input layer,
The lower stage corresponds to units 1 to 5 in the output layer. The horizontal direction represents time steps and corresponds to steps 1 to 30. The data given to each unit is displayed in area, and the maximum value is 1.0 and the minimum value is 0.0. That is, the input data obtained by density-shifting the analog image data and then normalizing the analog image data is treated as an area display here.

FIG. 13A shows learning data for causing the unit 1 of the output layer to output 1.0 when the input data to the unit 1 of the input layer is 0.6 or more. When data of 0.6 or more is continuously input to the unit 1, the data is trained to output 1.0 as a part of the character.

Similar to FIG. 13A, FIG. 13B is for making the unit 1 of the output layer output 1.0 when the input data to the units 1 and 2 of the input layer is 0.6 or more. This is learning data.

FIG. 13C shows learning data for causing the units 1 and 2 of the output layer to output 0.0 when the input data to the units 1 and 2 of the input layer is 0.4 or less.
When data of 0.4 or less is continuously input, it is learned that the input data is not a character and 0.0 is output.

In FIG. 13D, even if the input data to the input layer is 0.6 or more, if it is not continuous (3 steps or less), the input data is learned so as to output 0.0 as noise. 15 sets of data for outputting 1.0 for 0.6 input and 15 sets for outputting 1.0 for 1.0 input were also created. Furthermore, 12 sets of data that output 0.0 for noise were created.

Next, the result of the binarization processing according to this embodiment will be described. In this example, the RNN was trained with a total of 57 data as learning data. Learning is RIS
It was carried out by a workstation using a C chip, and the learning termination condition was a convergence error of 0.01 or less. As a result, the learning was completed 32,085 times and the learning time was about 3 hours and 14 minutes.

Since the size of the engraved character image is 100 × 100 dots, the image was divided into 20 in the size of 5 × 100 dots in the vertical direction. Twenty learned RNNs were prepared, the divided images were input in parallel to the respective RNNs, and the output result at that time was used as a binarized image.

FIG. 8 and FIG. 9 are examples of binarization results under the conditions of this embodiment.

In a high-contrast image, the contour of the character matches the contour of the character to be extracted as well as the binarization by the discriminant analysis method and RNN, and the binarization is performed with high accuracy. A small noise having an area of about 1 to 5 dots remains in the image by the discriminant analysis method, but the noise is removed by binarization by RNN.

In an image with a low contrast, noise is present in the character portion and background portion of the image by the discriminant analysis method so that the contour of the character cannot be confirmed, and the image is different from the contour of the character to be extracted. In the binarized image by RNN, the number of noises and the area are smaller than the image by the discriminant analysis method, and the contour of the character is closer to the contour of the character to be extracted.

In order to compare and examine the accuracy of binarization, the binarized image was recognized by a hierarchical neural network. Noise is removed and characters are cut out from the obtained binary image, and data for inputting to the hierarchical network is created from the cut out characters. In this report, as the data to be recognized by the network, a 25-dimensional vector is used, in which the extracted characters are divided into vertical and horizontal 5x5, and the area ratio of the character part in each mesh is normalized.
The number of units in each layer was 25 for the input layer, 10 for the intermediate layer, and 2 for the output layer. The network was trained with 4 letters each of 1 and 2 letters with no noise and a clear outline, for a total of 8 letters.

For recognition, a total of 150 characters, 69 characters 1 and 81 characters 2, were used. The correct answer was given when there was an output of 0.9 or more from the output layer.

FIG. 14 shows the recognition result. In the binarized image by RNN, two out of 150 characters could not be recognized, one was a character "1" and one was a character "2". These two characters cannot be recognized even in the binarized image by the discriminant analysis method. It can be said that among the 11 characters that could not be recognized in the binarized image by the discriminant analysis method, 9 characters became recognizable by performing binarization by RNN.

FIG. 15 shows two incorrect images in the binarized image by RNN. The letter "1" in the upper part of Fig. 15
It is considered that there was a large noise on the lower side of the character and could not be recognized due to the influence of this noise. But RN
In the binarized image by N, the noise itself remains, but the contour itself of the character 1 is close to the character to be extracted. The character "2" in the lower part of FIG. 15 has much noise on the contour of the character, and the contour of the character is different from the character to be extracted.

1 including an image with extremely low contrast
As a result of recognizing 50 characters, the number of correct answers of the binarized image by RNN was 148 and the recognition rate was 98.7%. Therefore, the binarization method according to the present embodiment can cope with a wide range of input conditions. It can be said that this is a binarization algorithm.

[0092]

Since the present invention is constructed and functions as described above, according to this, the threshold value selecting unit calculates the correction threshold value for binarizing the original image data and creates the input data. Section corrects the density level of the original image data based on the correction threshold value, the correction is performed regardless of the density of the analog image captured by the original image input section and regardless of the density distribution. The density level of the original image is corrected based on the threshold value. Therefore, the neural network processing unit
Since learning is performed using the correction threshold value as a reference value, it is possible to perform self-organization that is well matched to the feature of the recognition target. Therefore, binarization with higher accuracy can be performed as compared with the conventional method. As described above, it is possible to provide an unprecedented excellent image binarization device that can perform binarization with high accuracy without being affected by noise and contrast.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a configuration and an operation of a neural network processing unit shown in FIG.

FIG. 3 is a flowchart showing the operation of the embodiment shown in FIG.

FIG. 4 is an explanatory diagram showing an example of the relationship between the original image and the density distribution of one cross section thereof in the embodiment shown in FIG.

5A and 5B show a relationship between an original image and a density distribution in the embodiment shown in FIG. 1, FIG. 5A is a plan view showing the original image, and FIG. 5B is a case where a difference image is used. 5C is a density distribution map, FIG. 5C is a density distribution map showing the correction threshold value c, and FIG. 5D is a density distribution map in which the density level is corrected.

FIG. 6 is an explanatory diagram showing an operation of an input data creation unit in the embodiment shown in FIG.

7A and 7B are explanatory views showing the operation of the neural network processing unit in the embodiment of the present invention, FIG. 7A shows output data x1 for input data X1, and FIG. 7B shows output data for input data X2. x2 and FIG. 7C show the output data x3 with respect to the input data X3.

FIG. 8 is a plan view showing an example of a high-contrast original image and its binarized image. FIG. 8A is the original image and FIG.
8B is a binarized image according to the related art, and FIG. 8C is a binarized image according to the present invention.

9 is a plan view showing an example of a low-contrast original image and a binarized image thereof, FIG. 9 (A) being the original image and FIG.
9B is a binarized image according to the related art, and FIG. 9C is a binarized image according to the present invention.

FIG. 10 is a block diagram showing a configuration of a character recognition device to which the present invention has been applied.

11 is a diagram showing the relationship between the engraved characters and the density distribution in the embodiment shown in FIG. 10, and FIG. 11 (A) is a plan view of the original image “1” with high contrast and its density distribution diagram, FIG. 11B is a plan view of the original image “2” having a high contrast and its density distribution diagram, and FIG. 11C is a plan view of the original image “1” having a low contrast and its density distribution diagram, FIG. 11D is a plan view of the original image “2” having a low contrast and its density distribution diagram.

FIG. 12 is a flowchart showing an operation in the embodiment shown in FIG.

13 is an explanatory diagram showing a learning example of the neural network processing unit in the embodiment shown in FIG.
FIG. 13A is an explanatory diagram showing learning data for causing the unit 1 of the output layer to output 1.0 when the input data to the unit 1 of the input layer is 0.6 or more, and FIG. 13C is an explanatory diagram showing learning data that causes the units 1 and 2 of the output layer to output 1.0 when the input data to the units 1 and 2 of FIG. FIG. 13D is an explanatory diagram showing learning data for causing the units 1 and 2 of the output layer to output 0.0 when the input data to the input layers 2 and 3 is 0.4 or less, and FIG. It is explanatory drawing which shows the learning data which outputs 0.0 if data is not continuous even if it is 0.6 or more.

FIG. 14 is a chart showing a result of character recognition according to the embodiment shown in FIG.

15 is a plan view showing an incorrect image according to the embodiment shown in FIG.

[Explanation of symbols]

 10 original image input unit 12 threshold selection unit 14 input data creation unit 16 neural network processing unit 18 binarized image synthesis unit 20 input layer 22 intermediate layer 24 output layer b, A, B, C original image c for correction Threshold value X1 to Xn input data x1 to xn output data

Claims (1)

[Claims] 1. An original image input unit for capturing an image including a recognition processing target and converting the analog image into an original image having gradation, and calculating a correction threshold value for binarizing the original image. A threshold value selecting unit, an input data creating unit that creates input data by correcting the density level of the original image based on the correction threshold value, and normalizing the original image; An image binarization device, comprising: a neural network processing unit that binarizes input data created by the creating unit by neural network processing and outputs the binarized network.